"A solution in search for a problem" is a phrase used way to much, and almost always in the wrong way. Even in the article it says that it has been solving problems for over a year, it just complains that it isn't solving the biggest problems possible yet. It is remarkable how hard it is for people to extrapolate based on the trajectory. The author of this paper would have been talking about how pointless computers are if they were alive in the early 90s, and how they are just "a solution in search for a problem".
I am not a huge fan of generative AI, but even I can see it's potential (both for good and for harm). Today I found out about Suno in another thread on here, and tried it out. As a mid-millennial (1988) who grew up with CD players and still thinks MiniDiscs and ZIP discs are the coolest cartridge formats, aesthetically, that thing absolutely blows my mind.
We are like, 5 years into generative AI as a widely-available technology, and I can use it to generate entire songs on the fly based on just a couple sentences, complete with singing. I can use it to create logos and web graphics on my laptop in a matter of seconds, as I build a webpage. I can use it to help me build said webpage, also running locally on my laptop.
And it's still accelerating. 10 years from now, this stuff could be generating entire movies on-demand, running on a home media box.
Yeah, I absolutely agree. About a month ago, I would have said that Suno was clearly leading in AI music generation, but since then, Udio has definitely taken the lead. I can't imagine where things will be by the end of the year, let alone the end of the decade. This is why it's so crazy to me when people look at generative AI and act like it's no big deal and just a passing fad or whatever. They have no idea that there is a tsunami crashing down on us all and they always seem to be the ones that bill themselves as the weather experts who have it all figured out. Nobody knows the implications of this, but it definelty isn't an inconsequential tech.
The problem I see is mainly the divergence between hype and reality now, and a lack of a clear path forward.
Currently, AI is almost completely unable to work unsupervised. It fucks up constantly and is like a junior employee who sometimes shows up on acid. That's cool and all, but has relatively little practical use. However, I also don't see how this will improve over time. With computers or smartphones, you could see relatively early on, what the potential is and the progression was steady and could be somewhat reliably extrapolated. With AI that's not possible. We have no idea, if the current architectures could hit a wall tomorrow and don't improve anymore. It could become an asymptotic process, where we need massive increases for marginal gains.
Those two things combined mean, we currently only have toys, and we don't know if these will turn into tools anytime soon.
Yeah, it's trajectory thing. Most people see the one-shot responses of something like the chatgpt's current web interface on openai's website and they think that's where we are at. It isn't though, the cutting edge of just what is currently openly available to people is things like CrewAI or Autogen using agents powered by things like Claude Opus or Llama 3, and maybe the latest gpt4 update.
When you use agents you don't have to baby every response, the agents can run code, test code, check latest information on the internet, and more. This way you can give a complex instruction, let it run and come back to a finished product.
I say it is a trajectory thing because when you compare what was cutting-edge just 1 year ago, basically one-shot gpt3.5 to an agent network with today's latest models, the difference is stark, and when you go a couple years before that to gpt2, it is way beyond stark. When you go a step further and realise that there is lots of custom hardware being built(basically llm ASICs-traditionally a ~10,000x speedup over general use gpus), you can see that soon having instant agent based responses will be the norm.
All this compounds when you consider that we have not hit a plateau and that we are still seeing that better datasets, and more compute, are still producing better models. Not to mention that other architectures, like state-based Mamba, are making remarkable achievements with very little compute so far. We have no idea how powerful thinks like Mamba would be if they were given the datasets and training that the current popular models are being given.
Extrapolation is, like, one notch above guessing, though. It's not wrong, exactly, but I'm not convinced failing to do it is an error in every context.
Mostly, you're right, this article makes it's argument by openly ignoring all the applications it has found. But anything where "hallucination" would be a problem might need a fundamentally different technology.
I think without anything akin to extrapolation, we just need to wait and see what the future holds. In my view, most people are almost certainly going to be hit up side the head in the not to distant future. Many people haven't even considered what a world might be like where pretty much all the jobs that people are doing now are easily automated. It is almost like instead of considering this, they are just clinging to some idea that the 100-meter wave hanging above us couldn't possibly crash down.
My issue with generative AI is not that it doesn't have uses, but that it seems to me that the vast majority of those uses are nefarious.
As far as I can tell, it has the most potential for:
-
Creating sock puppet accounts on social media to sway public opinion
-
Make fake media/ identity theft
-
Plagarize various art mediums and meld them together enough to make attribution difficult
Other positive use cases like summarization or reformatting seem to pale in comparison to the potential negative effects of the bad use cases. There are many marginal use cases like coding or law where you may save some time but the review required is likely not that much different than the time it would take for a good programmer or lawyer to just write it.
Most positive use cases are agent-based and the average user doesn't have access to good agent-based systems yet because it requires a bit of willingness to do some "coding". This will soon not be the case though. I can give my crew of AI agents a mission, for example, "find all the papers on baby owl vocalizations and make 10 different charts of the frequency range relative to their average size after each of their first 10 weeks of life", and come back an hour later and have something that would have been 100 hours for a grad student just last year. Right now I have to wait an hour or so, soon it will be instant.
The real usefulness of these agents today is enormous, it is just outside of the view of many average people because their normal lives don't require this kind of power.
You forgot porn.
Edit: Actually, in the article it mentions coding assistents and various interfaces. Not to mention the plagiarism thing is a misunderstanding. I'm not sure why I decided to jump on the jerk there, I disagree with you.
-
Agreed. I mean yeah, image generators are still very limited (or at least, difficult to use in an advanced, targeted way), but there's a new research paper out every day detailing new techniques. None of the criticisms of Midjourney or Stable Diffusion today are likely to remain valid in a year or even six months. And they're already highly useful for certain tasks.
Same with LLMs, only we've already reached the point where they are good enough for almost anything if you care to write a good application around them. The problem with LLMs at this point is marketing; people expect them to be magic and are disappointed when they don't live up their expectations. They're not magic but they are extremely useful. Just please, for the love of god, do not treat them as information repositories...
Isn't the "trajectory" that these systems are incredibly unsustainable both economically and environmentally? I'd hope that a machine that uses a few thousand homes' worth of energy to answer a single query would be more useful than "can generate boilerplate code for me" or whatever.
I think there may be some confusion about how much energy it takes to respond to a single query or generate boilerplate code. I can run Llama 3 on my computer and it can do those things no problem. My computer would use about 6kWh if I ran it for 24 hours, a person in comparison takes about half of that. If my computer spends 4 hours answering queries and making code then it would take 1kWh, and that would be a whole lot of code and answers. The whole thing about powering a small town is a one-time process when the model is made, so to determine if that it worth it or not it needs to be distributed over everyone who ends up using the model that is produced. The math for that would be a bit trickier.
When compared to the amount of energy it would take to produce a group of people that can do question answering and code writing, I'm very certain that the ai model method is considerably less. Hopefully, we don't start making our decision about which one to produce based on energy efficiency. We may, though, if the people that choose the fate of the masses sees us like livestock, then we may end up having our numbers reduced in the name of efficiency. When cars were invented, horses didn't end up all living in paradise. There were just a whole lot less of them around.
That's going to change in the future with NPUs (neural processing units). They're already being bundled with both regular CPUs (such as the Ryzen 8000 series) and mobile SoCs (such as the Snapdragon 8 Gen 3). The NPU included with the the SD8Gen3 for instance can run models like Llama 2 - something an average desktop would normally struggle with. Now this is only the 7B model mind you, so it's a far cry from more powerful models like the 70B, but this will only improve in the future. Over the next few years, NPUs - and applications that take advantage of them - will be a completely normal thing, and it won't require a household's worth of energy. I mean, we're already seeing various applications of it, eg in smartphone cameras, photo editing apps, digital assistants etc. The next would be I guess autocorrect and word prediction, and I for one can't wait to ditch our current, crappy markov keyboards.
I enjoy image generation AI for it's ability to turn that really specific picture you have in your head into something that you can show to others within a matter of seconds.
Weird, I use it all the time. Even starting to use it for work to save a ton of time on simple, time-consuming work.
It's use in hybrid search CRAG and discussion systems as a human in the loop augmentation is quite valuable. It saves analyst time and streamlines further review. In a sufficiently adversarial system where multiple models are vetting the responses for sanity with back checks, it's quite performant, having decent recall and accuracy.
It clearly isn't!
A) Man has dreamt of Artificial Intelligence for decades now, often times very much realizing the capabilities (and dangers) of such technology B) AI in its current form already support business, hobbies, creative work etc. The traffic and processing power needed is constantly rising.
I feel with a such a bold (and just incorrect) statement the article cannot be much worth to read.
